GLM-5.3-Flash (Ox Alpha) Launch

Z.ai officially announced GLM-5.3-Flash (previously previewed as Ox Alpha), a 320B-total/18B-active MoE natively multimodal model with a 1M-token context window, released under MIT License and running entirely on Chinese AI chips. - Performance: DeepSWE v1.1 63.4 (GLM-5.2 46.2, Opus 4.8 58.0); Terminal Bench 2.1 84.3 (Opus 4.8 85.0); AutomationBench 48.8 (GLM-5.2 26.2, Opus 4.8 41.0); Toolathlon Verified 78.4, highest recorded; internal coding eval 29.0 vs. Opus 4.8 29.5. - Cost: Artificial Analysis Intelligence Index 57, on par with Opus 4.8; discounted task cost $0.045, about 1/10 of previous models at this intelligence level; limited half-price offer making it 1/40 of Opus 4.8. - Weaknesses: NL2Repo 56.3 (Opus 4.8 69.7); visual/video benchmarks (BabyVision, MVBench, MMVU) lag behind Gemini 3.7 Flash. - Natively multimodal: model actively observes rendered output and iterates for front-end, gaming, Blender 3D, browser/computer use; training uses visual-judgment data synthesis and RL with environment feedback; can end-to-end deliver PPTX/PDF/DOCX/XLSX. - Chinese chip deployment: anonymous ox-alpha ran entirely on domestic accelerators, became most popular model on OpenCode/OpenRouter; inference engine uses SGLang with tensor parallelism, ReplaySSM, W8A8 quantization, mixed cache quantization, layer split; cluster employs EPD separation, scheduling tens of thousands of domestic GPUs; 3x end-to-end inference speed improvement, per-token cost comparable to NVIDIA GPUs; infrastructure agent (powered by GLM-5.3) assisted engine optimization. - Availability: API, coding plan, chat, weights open-source; free beta started last Friday, formal version next day; global daily supply >100T entirely served by Chinese chips. - Market adoption: within 3 days of preview, Ox Alpha usage surpassed DeepSeek V4 Flash on OpenCode.
Key insights
Benchmark Performance Nearly Matches Opus 4.8 in Coding and Agent Tasks
GLM-5.3-Flash delivers DeepSWE v1.1 score of 63.4 (Opus 4.8 58.0), Terminal Bench 2.1 84.3 (Opus 4.8 85.0), AutomationBench 48.8 (Opus 4.8 41.0), and Toolathlon Verified 78.4, leading the leaderboard. An internal coding benchmark run on Claude Code gave 29.0, within 0.5 of Opus 4.8. These results position it close to Opus 4.8 while the model's active parameters are only 18B, underscoring architectural efficiency.
Ultra-Low Cost Redefines Price-Performance Ratio
With an Artificial Analysis Intelligence Index of 57 and a discounted task cost of $0.045, GLM-5.3-Flash offers about one-tenth the cost of previously comparable models. A limited half-price promotion brings its cost to 1/40 of Opus 4.8, creating a new cost-effectiveness kill line. The model's 18B active parameters from a 320B MoE architecture drive this price advantage, making high-level AI dramatically cheaper.
Native Multimodal Vision Loops Enable Real-World Interaction
GLM-5.3-Flash natively embeds vision in its coding loop, allowing it to observe rendered output, interpret feedback, and iteratively correct itself across front-end development, gaming, Blender 3D scenes, and browser/computer use. The training pipeline includes a self-visual-judgment data synthesis system and reinforcement learning with environment feedback. This capability extends to document understanding, enabling the model to produce end-to-end PPTX, PDF, DOCX, and XLSX files.
Chinese Chip Infrastructure Powers Entire Service at Scale
Before its official release, the model ran anonymously on OpenCode/OpenRouter entirely on Chinese-made accelerators, handling over 100T daily supply and becoming the week's most popular model. The inference engine uses SGLang with intra-node tensor parallelism, ReplaySSM, W8A8 quantization, mixed-precision cache quantization, and layer splitting. The cluster architecture separates encode, prefill, and decode stages, scheduling tens of thousands of domestic chips. End-to-end inference performance improved 3× versus the baseline, with per-token cost reaching parity with mainstream NVIDIA GPUs. Notably, the inference engine optimization was partly assisted by infrastructure agents powered by GLM-5.3 itself.
Weaknesses Remain in Code Repository Generation and Video Benchmarks
Despite strong overall performance, GLM-5.3-Flash scores only 56.3 on NL2Repo, substantially behind Opus 4.8's 69.7, indicating weaker code-repository generation. It also underperforms Gemini 3.7 Flash on several vision and video benchmarks, including BabyVision, MVBench, and MMVU, highlighting gaps in visual reasoning tasks.
Rapid Market Adoption Surpasses DeepSeek V4 Flash in Days
Usage data shared after the Ox Alpha preview showed that within 3 days, the model's traffic on OpenCode completely overtook DeepSeek V4 Flash, represented by pink and purple bars. This rapid shift suggests that when a new model offers better performance, the developer community switches quickly, and that any lab can win market share by releasing the best model.